INRIA a CCSD electronic archive server
Not a member yet
122212 research outputs found
Sort by
Moderate-length lifted quantum Tanner codes
No full textWe introduce new families of quantum Tanner codes, a class of quantum codes that first appeared in the work of Leverrier and Zémor [LZ22]. These codes are built from two classical Tanner codes, for which the underlying graphs are extracted from coverings of 2D geometrical complexes, and the local linear codes are tensor-products of cyclic or double-circulant linear codes. The advantage of code lifting is that, for any lift of odd index t of an [[n, k, d]]-code, we can adapt the study of the transfer homomorphism arising in cellular homology to describe symmetries of its logical operators and to establish that its dimension is lower bounded by k, and its distance is upper bounded by t • d. Moreover, when the dimension of the lifted code is equal to k, its distance is lower bounded by d. These parameter bounds also apply to the previous methods of code lifting [Gue25]. Finally, We present several explicit families, and identify instances of moderate length quantum codes which are degenerate, have low check weight, and whose distance surpasses the square root of the code length. Among them, we report the existence of a [[96, 2, 12]]-code whose distance growth saturates our bound, and for which half of the checks are of weight 8 and the other half of weight 4
Enhanced 3D Object Detection via Diverse Feature Representations of 4D Radar Tensor
No full textInternational audienceRecent advances in automotive fourdimensional (4D) Radar have enabled access to raw 4D Radar Tensor (4DRT), offering richer spatial and Doppler information than conventional point clouds. While most existing methods rely on heavily pre-processed, sparse Radar data, recent attempts to directly leverage raw 4DRT incur high computational costs and limited scalability. To address these limitations, we propose a novel three-dimensional (3D) object detection framework that explicitly addresses the representation-level variability inherent in 4D Radar while preserving efficiency. Rather than assuming a single optimal Radar representation, our method introduces a Radar-centric multi-teacher knowledge distillation (KD) framework, where multiple teacher models are trained on point clouds derived from diverse 4DRT pre-processing techniques, each capturing complementary signal characteristics. These teacher representations are fused via a dedicated aggregation module and distilled into a lightweight student model that operates solely on sparse Radar inputs. Experimental results on the K-Radar dataset demonstrate that our framework achieves improvements of 7.3% in AP 3D and 9.5% in AP BEV over the baseline RTNH model when using extremely sparse inputs. Furthermore, it attains comparable performance to denser-input baselines while significantly reducing the input data size by about 90 times, confirming the scalability and efficiency of our approach
New Metrics of Event-Related (De)Synchronization Temporal Variability Explain Motor Imagery-based BCI Performance
No full textMotor Imagery-based (MI) Brain-Computer Interface (BCI) detect imagined limb movements from ElectroEncephaloGraphy (EEG) to translate them into commands for various applications. They do so by analyzing sensorimotor EEG rhythms, typically event-related (de)synchronization (ERD/S) over the motor cortex. Despite MI task intuitiveness and their many BCI applications, not all users achieve sufficient MI classification accuracy, notably due to large intraand inter-user variability in ERD/S. Understanding this variability is thus crucial for finding ways to enhance BCI classification performance, but BCI variability metrics are lacking. Therefore, this paper proposes two new ERD/S variability metrics and studies, on a large MI-BCI dataset (N = 85 users), how these and two existing metrics can explain BCI performance.Results show that temporal variability of ERD/S—both within and across trials—negatively correlates (r = −0.28 to −0.34) with BCI performance in the within-user scenario (with a user-specific classifier). In the cross-users scenario (with a generic cross-user classifier), test users variability metrics, including ERD/S temporal and amplitude variability, were negatively correlated with performance (r = −0.30 to −0.39). These findings demonstrate the value of metrics to quantify ERD/S variability. They may also guide future design strategies for BCI user training or machine learning
2nd Workshop on Scheduling Variable CapacityResources for Sustainability
No full textThis is the report for the second workshop on Scheduling Variable Capacity Resources for Sustainability. The first workshop in 2023 nucleated a research community focused on compute scheduling in the new age of renewable power generation -- where variation in weather and solar radiation drives variation in compute capacity (opportunity). The goal was to mobilize a combination of the scheduling community, cloud resource management community, and new leaders creating adaptive datacenters. These adaptive grid loads could then accelerate the use of clean renewable energy and thereby power grid decarbonization. This first workshop posited 1) variable capacity, flexible compute platforms; 2) dynamic grids and datacenter-grid decoupling; 3) adaptive scheduling approaches to accommodate these varying platforms and respond to grid demands; and 4) a growing social and government demand for carbon progress in datacenter systems. In September 2025, all of these technical changes come to reality, at large scale because of the AI fueled explosion of datacenter power demand and grid stress.However, in September 2025, the world of variable capacity scheduling is also shaped by two radical changes. First, a geo-political shock. While the dangers of climate change continue to grow, the turn of the US government's policies away from renewables, even promoting fossil fuels, has created a global turmoil. Second, a workload shock. Over the past two years, the AI revolution dramatically accelerated the growth of electric power consumption by computing -- in North America, Europe, Asia (notably China), and the Middle East. In short, the trends that motivated the workshop have only increased in importance and urgency.At the second workshop, we assessed progress made in 2.5 years on the key problems of variable capacity scheduling, and also trends, challenges, and opportunities:- Platforms are rapidly increasing in dynamics, with growing adoption of dynamic power management and adaptive scheduling to meet demand-response and other grid needs. There is a growing two-way relationship between datacenters and grids;opportunities abound.- Workloads become more malleable, delay flexible, or even acceptably approximate. AI training workloads are a major example, with hyper-parameter optimization and checkpoints enabling significant flexibility. - The scheduling algorithms and metrics are evolving rapidly, reflecting the interplay between flexible workloads and dynamically varying platforms. New metrics are emerging to capture both performance and non-performance attributes (e.g. carbon emissions), and new notions of progress (model improvement).- Societal concerns about the sustainability of computing have exploded into the public eye with growing resistance and protest against datacenters. Criticism of perpetual growth gives rise to questions about new models of sufficiency, as well as growing awareness, responsibilities, and action.The talks and position papers of the participants are available at https://people.cs.uchicago.edu/~aachien/workshops/varsched25/
Learning to Act Greedily: Polymatroid Semi-Bandits
No full textInternational audienceMany important optimization problems, such as the minimum spanning tree and minimum-cost flow, can be solved optimally by a greedy method. In this work, we study a learning variant of these problems, where the model of the problem is unknown and has to be learned by interacting repeatedly with the environment in the bandit setting. We formalize our learning problem quite generally, as learning how to maximize an unknown modular function on a known polymatroid. We propose a computationally efficient algorithm for solving our problem and bound its expected cumulative regret. Our gap-dependent upper bound is tight up to a constant and our gap-free upper bound is tight up to polylogarithmic factors. Finally, we evaluate our method on three problems and demonstrate that it is practical
Fragments d'Optimisation Différentiable - Théories et Algorithmes
No full textMasterLecture Notes (in French) of optimization courses given at ENSTA (Paris, next Saclay), ENSAE (Paris) and at the universities Paris I, Paris VI and Paris Saclay (1019 pages).Syllabus d’enseignements délivrés à l’ENSTA (Paris, puis Saclay), à l’ENSAE (Paris) et aux universités Paris I, Paris VI et Paris Saclay (1019 pages)
Multiscale analysis of a kinetic equation for mechanotaxis
No full textWe present a new kinetic equation for cell migration driven by mechanical interactions with the substrate, an effect not previously captured in kinetic models, and essential for explaining observed collective behaviors such as those in bacterial colonies. The model introduces an acceleration term that accounts for the dynamics of motile cells undergoing mechanotaxis, where extracellular signals modulate the forces arising from cell-substrate interactions. From this formulation, we derive a family of macroscopic limit equations and analyze their principal properties. In particular, we examine linear stability and pattern formation ability through theoretical analysis, supported by numerical simulations
ComposeAnything: Composite Object Priors for Text-to-Image Generation
No full textInternational audienceCurrent text-to-image models struggle to generate scenes with many objects and complex relations. Training-time solutions such as layout conditioning or reinforcement learning improve compositional accuracy but often degrade image quality and realism by enforcing rigid constraints. To address this limitation, we introduce ComposeAnything, an inference-only framework that injects a structured composite object prior directly into the diffusion process. Rather than starting from random latent noises or performing expensive noise optimization, we construct a single 2.5D composite prior encoding strong object appearance, counts, sizes, and coarse depth-aware placement, and use it to initialize and guide one diffusion trajectory. This explicit prior is interpretable and editable in image space, enabling human-in-the-loop refinement by simply adjusting the composite. Our training-free, backbone-agnostic method improves compositional consistency on T2I-CompBench and NSR-1K benchmarks, particularly for complex prompts, while maintaining high visual quality compared to both training-based baselines and other inference-time methods
Multi-criteria and multi-stage environmental study of Pl@ntnet service for the year 2024
No full textIn this study, we focus our investigation on Pl@ntNet, a citizen science platform, which re- lies on Artificial Intelligence (AI) models to identify plant species. Pl@ntNet provides a large-scale infrastructure supporting millions of users in over 200 countries. At this stage of deployment, and with years of experience developing the platform, Pl@ntNet is committed to understanding the environmental impacts of its identification service and contributing to the search for reduction opportunities. Our investigations assess the associated environmental impacts of Pl@ntNet for the year 2024. We based our approach on multi-criteria LCA, considering multiple impact type and the different life-cycle phases
Toward Real-Time RAN Observability in Open-Source 5G Systems
No full textInternational audienceDiagnosing performance issues in 5G and future disaggregated networks remains challenging due to the complexity of the protocol stack and the multitude of interdependent metrics. This is especially true in experimental environments using opensource 5G software, where logs are often verbose, fragmented, and difficult to interpret. Real-time visual observability, especially at the Radio Access Network (RAN) level, is therefore essential for effective troubleshooting. However, tools like 5GC-Observer and Monarch provide only partial support, focusing mainly on the core network and lacking visibility into the RAN, where many performance bottlenecks originate. To tackle this, we present a Prometheus-compatible telemetry pipeline for the OpenAirInterface (OAI) gNB that extracts RAN metrics, adjusts acquisition intervals via a custom FlexRIC model, and visualizes the data in real time with Grafana. The pipeline is also compatible with SRSRAN through a lightweight Telegraf-Prometheus integration, extending its use to multiple open-source stacks. We integrate our solution into the Monarch monitoring architecture for cloud-native 5G, enabling end-to-end observability. An Ansible-based automation simplifies testbed setup and ensures reproducible experimentation. Validation on a realistic testbed shows real-time RAN metrics exposure with negligible overhead